the simulation of urban-scale evacuation scenarios with …gala.gre.ac.uk/id/eprint/18045/1/18045...
TRANSCRIPT
The Simulation of Urban-Scale Evacuation Scenarios with application
to the Swinley Forest Fire Anand Veeraswamy1, Edwin R Galea1, Lazaros Filippidis1, Peter J Lawrence1, Simo Haasanen1,
Robert J Gazzard2 and Thomas E L Smith3 1 Fire Safety Engineering Group
University of Greenwich
London SE10 9LS
United Kingdom
2 Forestry Commission England
Bucks Horn Oak
Surrey GU10 4LS
United Kingdom 3 Department of Geography
King’s College London
London WC2R 2LS
United Kingdom
Abstract
Forest fires are an annual occurrence in many parts of the world forcing large-scale
evacuation. The frequent and growing occurrence of these events makes it necessary to
develop appropriate evacuation plans for areas that are susceptible to forest fires. The
buildingEXODUS evacuation model has been extended to model large-scale urban
evacuations by including the road network and open spaces (e.g. parks, green spaces
and town squares) along with buildings. The evacuation simulation results have been
coupled with the results of a forest fire spread model and applied to the Swinley forest
fire which occurred in Berkshire, UK in May 2011. Four evacuation procedures
differing in the routes taken by the pedestrians were evaluated providing key evacuation
statistics such as time to reach the assembly location, the distance travelled, congestion
experienced by the agents and the safety margins associated with using each evacuation
route. A key finding of this work is the importance of formulating evacuation
procedures that identifies the threatened population, provides timely evacuation notice,
identifies appropriate routes that maintains a safe distance from the hazard front thereby
maximising safety margins even at the cost of taking longer evacuation routes.
Evacuation simulation offers a means of achieving these goals.
KEYWORDS: urban evacuation, evacuation simulation, wildfire, forest fire, disaster management, GIS
1 Introduction
Forest fires (also referred to as wildfires and bushfires) are an annual occurrence in
many parts of the world causing massive damage to the natural and built environments,
endangering lives and forcing the large-scale evacuation of entire residential areas and
business and industrial facilities. Recent examples include the Fort McMurray fires
(The Globe and Mail, 2016) in Canada (May 2016) which extended over an area of
590,000 hectares of Alberta involving over 4,000 fire fighters and other emergency
personnel and resulted in the loss of 2,000 structures, the forced evacuation of 88,000
people and damage estimated at $CAN3.6 billion (Regional Municipality of Wood
Buffalo, 2016); the San Bernardino County Blue Cut fires in California (August 2016)
involving more than 2,600 fire fighters, resulting in the loss of over 100 family
residences (Incident Information System, 2016) and requiring mandatory evacuation
calls to over 34,000 homes with more than 82,000 people, ranging from a ski resort to
the entire town of Phelan with more than 14,000 residents (Chicago Tribune, 2016);
and the Hafia Israel fires (BBC News, 2016), resulting in the evacuation of some 80,000
people from homes, schools, kindergartens, universities, a hospital for the aged and
prisons (BBC News, 2016 and The New York Times, 2016), police and emergency
services went door to door to alert people on the need to evacuate and the affected
population evacuated on foot and by car (The New York Times, 2016).
One of the world’s worst wildfires in living memory are the Black Saturday bushfires
of Victoria Australia (February 2009) which claimed the lives of 173 people, burnt
430,000 hectares of land, 2000 properties and 61 businesses, with an estimated cost of
$AUS4 billion (Teague et al. 2009). Rather than a single fire there were over 47 fires
scattered across the State that had the potential to become severe, five of which claimed
people’s lives – the most costly was the Kilmore East fire that claimed 119 lives. The
Country Fire Authority (CFA) engaged approximately 12,000 CFA fire fighters and
over 1000 operational vehicles to fight the fires with the support of an additional 1000
fire fighters from the Department of Sustainability and the Environment (DSE). One
of the issues to come out of the Royal Commission held to investigate the response to
the fires concerned the advice given to the community on whether or not they should
evacuate, the so-called ‘stay or go’ policy, more correctly stated as, ‘Prepare, Stay and
Defend or Leave Early’. The Commission concluded that the authorities had placed
insufficient emphasis on the risks of staying and defending ones property and that the
safest option is always to leave early. The Commission recommended that, ‘… the
State introduce a comprehensive approach to evacuation so that this option is planned,
considered and implemented when it is likely to offer a higher level of protection than
other contingency options ……’ (Recommendation 5) (Teague et al., 2010).
The threat posed by wildfires is not restricted to countries with hot climates. A large
number of wildfires occur in Great Britain every year. In the past four years there have
been on average 45,000 wildfires each year attended by the fire and rescue services in
Great Britain (Crowhurst, 2015). Furthermore, warmer and drier conditions, and more
frequent and longer-lasting heatwaves also raise the risk of wildfires. This risk is
compounded by the UK’s high population density, which means that fires are more
likely to encroach into urban environments, posing a significant threat to life. Forest
fires can spread rapidly affecting built-up areas in its path at different times depending
on the speed and direction of the fire spread, which in turn is dependent on a number
of factors such as the wind speed and direction, nature of terrain and available fuel in
the path of the fire (BBC News, 2016, Chicago Tribune 2016, Regional Municipality
of Wood Buffalo, 2016, Teague et al., 2009, Teague et al., 2010 and The New York
Times, 2016).
It is clear that wildfire often requires the evacuation of large numbers of people from
large areas and potentially over great distances. The frequent and growing occurrence
of these events suggests that it is necessary to develop appropriate evacuation plans for
areas that are susceptible to wildfires.
This paper provides an overview of the development of the EXODUS urban-scale
evacuation modelling system (Chooramun et al., 2012, Galea et al., 2008, Galea et al.,
2011, Pretorius et al., 2013 and Siddiqui and Gwynne, 2012), and describes the
application of the software to a case study involving the Swinley Forest fire that
occurred near Bracknell in Berkshire, UK on May 02, 2011. As part of this work, the
paper also describes the loose coupling of the EXODUS software to the Prometheus
wildfire simulation tool (Tymstra et al., 2010). It is suggested that the developed
methodology could be utilised by authorities to formulate evacuation procedures before
an incident takes place based on historic data (e.g. past wildfires, floods, earthquakes,
etc.) and also has scope to be applied during incidents by providing first responders
with key evacuation related information.
2 Literature Review
When planning an evacuation for a community threatened by wildfire, it is essential for
civil protection authorities to know what areas will be affected due to a fire, when the
areas will become untenable, how long it will take occupants in these regions to
evacuate and how long it will take occupants to reach a designated place of safety. To
this end, fire simulation tools such as Prometheus (Tymstra et al., 2010) and Phoenix
(Tolhurst et al., 2008) have been developed that attempt to predict the spread of the
wildfire. These models take into consideration the nature of the available fuel, the
terrain and the atmospheric conditions and provide an estimation of the direction and
rate of spread of the fire front.
While this type of information is necessary to assess the level of risk associated with
each region it is not sufficient to optimally determine how and when to warn the
population appropriately, how to allocate available resources (first responders, fire
fighters and staff at assembly locations) and most importantly to formulate an effective
evacuation strategy which involves safe routes that occupants can take to the designated
assembly locations. To do this also requires an ability to predict the evacuation
movement and behaviour of city sized populations of people measuring in the tens of
thousands if not hundreds of thousands.
Several evacuation models have been developed that attempt to simulate large scale
evacuation situations resulting from floods (Durst et al., 2014) and earthquakes
(D’Orazio et al., 2014). MATSIM, a transport simulation tool has been used to perform
pedestrian evacuation simulation at city scale involving over a million agents (Lämmel,
2008). The road network is imported from OpenStreetMaps (Neis et al., 2011) and
represented in MATSIM as a directed graph with the streets being represented as links
or arcs connecting the vertices or nodes, which represent a point on the earth’s surface.
Network change events such as altering the state of a link from traversable to non-
traversable occurs dynamically as the simulation runs. Initially all agents utilise a
shortest path to the refuge areas. However, the Nash equilibrium (Osborne and
Rubinstein, 1994) approach is employed where agents attempt to find a route that is
optimal in terms of the time required to reach the destination in each successive iteration
taking into account congestion levels encountered on the evacuation routes. Over many
iterations this approach provides an estimation of the overall shortest evacuation time
which could represent an evacuation where agents have been trained or guided so as to
follow a pre-determined procedure. This approach has been used to model the
pedestrian evacuation simulation of the entire city of Padang, Indonesia involving
320,000 agents in response to a tsunami (Lämmel et al., 2010) and a part of Zurich
involving 165,571 agents in response to a flood (Lämmel, 2008). A similar approach
has been used to model large-scale multi-modal (i.e. involving pedestrians, buses,
railway and cars) evacuation simulation of Hamburg-Wilhelmsburg, Germany
consisting of 50,601 inhabitants in response to a flood (Klüpfel, 2014). Travel mode
change options were also implemented allowing for example agents to walk to a bus
stop, take a bus to the train station and take a train to a safe location.
The PedGo simulation tool has been used by (Klüpfel, 2014) to perform pre-event and
post-event pedestrian simulations of a popular event that attracted a very large number
of visitors (250,000) in Burgdorf, Switzerland. The actual pedestrian traffic flow and
congestion observed during the event was significantly different (less congestion was
observed than predicted) from the pre-event simulations mainly because of an
underestimation of response times and overestimation of the number of people. When
these factors were modified accordingly the post-event simulations provided a good
agreement with the observations during the actual event. This illustrates the intricacies
in large scale evacuations - small deviations from assumptions of initial conditions can
lead to large differences in the simulation results when compared to the actual event.
The simulation of large-scale evacuation requires the inclusion of a number of human
behaviour aspects into the modelling process. According to (Osaragi et al., 2013) the
following factors are key to model large scale evacuations following an earthquake:
spatiotemporal distribution of people at the time of the earthquake, time at which people
start evacuating, decision to head to a temporary refuge location or to the official refuge
location and routes that people choose to use (familiar versus safe routes avoiding
affected areas). A study of human behaviour during earthquakes, both within structures
and outside, (Bernardini et al., 2014) has identified the following behaviours;
attraction to safe areas, herding behaviour, attraction between members of the same
group, keeping a safe distance from buildings and group formation. They also
determined the average speeds and distances between members of the same evacuating
group and represented these behaviours in a simulation tool by modifying the social
force model (Helbing and Johansson, 2013).
A city scale evacuation model utilising a potential map system for navigation has been
developed (Nishino et al., 2011) where agents travel towards refuge locations taking
paths that they perceive to be safe while staying away from hazards such as fire. A flood
simulation model was coupled to an evacuation simulation model (Mordvintsev et al.,
2014) where agents could assume one of several states such as idle, running, safe and
drowned. Agents followed a potential map avoiding static obstacles as well as non-
traversable flooded areas to reach safe refuge areas.
While several wildfire evacuation models have been developed (Cova and Johnson,
2002 and Singh et al., 2016), most of these are vehicular models (Cova and Johnson,
2002 and Singh et al., 2016) thus ignoring pedestrian movement. Furthermore, few if
any of the wildfire evacuation models are coupled with fire simulation tools and when
this is done it is generally for visualisation purposes only as in the Singh et al model
(Singh et al., 2016) which utilises output from the Phoenix fire simulation tool (Tolhurst
et al. 2008). Coupling agent based urban scale evacuation models to wildfire simulation
tools enables the identification of, the available time for safe evacuation, which routes
become untenable and at what time, population exposure doses to harmful fire products
during evacuation and safety margins associated with alternative evacuation strategies.
Thus while evacuation models have been developed that attempt to simulate large scale
evacuation situations most of these are vehicular evacuation models or adapt vehicle
simulation tools to model pedestrian evacuation utilising only the vehicular network
thus ignoring areas/routes that could be utilised by pedestrians and utilise a macroscopic
model for representing space and agent movement. Furthermore, current urban scale
agent based evacuation modelling tools are generally not integrated with disaster
management systems. As a result, disaster managers do not have an easy means to test
evacuation procedures during pre-incident planning or receive real time support on
optimal evacuation procedures to adopt during an ongoing incident. Integrating
evacuation simulation tools with disaster management systems to produce a Common
Operating Picture (COP) provides valuable evacuation decision support functionality
for disaster managers to assist in the co-ordination and management of disasters.
Large scale disasters bring together diverse organizations which produce large
quantities of heterogeneous data. This makes it vital to utilise Information and
Communications Technology (ICT) to co-ordinate disaster management thereby
making it more effective. Disaster management systems have various components such
as organisation registry, missing persons registry, shelter registry (camps, hospitals,
schools), request management system, inventory management system, etc (Careem et
al., 2007 and Van de Walle and Turoff, 2008). Current disaster management systems
such as Sahana (Careem et al., 2006 and Currion et al., 2007) typically utilise a
Geographic Information System (GIS) to capture, store, analyse and manage
geographically referenced data to provide a coherent system for incident commanders.
They also rely greatly on open standards to enable communication between diverse
systems. However, most currently available disaster management systems do not
presently integrate external simulation tools. In order for external simulation tools to be
integrated with disaster management systems the simulation tools need to be able to
read and produce outputs using open standards.
While the CEMPS (Configurable Emergency Management and Planning Simulator)
simulator (Silva and Eglese, 2000) comes close to addressing this problem by coupling
a traffic simulation model with ESRI’s GIS-ARC/INFO, nothing currently exists that
couples agent based evacuation models with incident management systems.
The approach adopted in this paper differs from previous work by providing a loose
coupling between the evacuation simulation tool and disaster management system with
the focus being on pedestrian rather than vehicle traffic simulation in response to large
scale disasters. The agent based evacuation simulation tool buildingEXODUS while
primarily used for building evacuation applications (Chooramun et al., 2012, Galea et
al., 2008 and Galea et al., 2011) has also been used to simulate large crowds in external
environments (Pretorius et al., 2013 and Siddiqui and Gwynne, 2012). As part of the
EU FP7 project IDIRA (Fire Safety Engineering Group, 2015b), the buildingEXODUS
software was adapted to model large scale external spaces involving not simply the road
network as current large scale evacuation models do (Lämmel et al., 2008, Lämmel et
al., 2010) but also involving other objects utilised by pedestrians such as buildings,
open spaces, paths, etc. It also utilised open GIS standards such as shapefiles (Xia and
Wei, 2008), Web Map Service (Li et al., 2010) and Web Feature Service (Peng and
Zhang, 2004) and Free and Open Source Software (FOSS) such as such as PostGIS
spatial extensions (Obe and Hsu, 2011) for PostgreSQL database (Drake and Worsley,
2002), OpenLayers (Perez, 2012) a JavaScript library for displaying map data in web
browsers to develop a user friendly GUI that the strategic (command and control
centres) and tactical (on-scene) commanders could utilise to specify input data for the
EXODUS simulation such as: areas to evacuate, refuge locations, non-traversable areas
and distribution and attributes of pedestrian population. The GUI also enables
commanders to run simulations and analyse the results in order to determine the
efficiency of the various evacuation options available. In this way the EXODUS
software was integrated within the web based GIS disaster management system (COP)
developed as part of the IDIRA project. The developed system was found to be a
valuable decision support tool for the people/organisations involved in disaster
management.
3 The EXODUS Large Scale Evacuation System
The EXODUS large scale evacuation modelling system is an agent based
evacuation model capable of simulating the evacuation of large populations – measured
in the hundreds of thousands - in large scale environments – measuring many square
kilometres. It is based on the buildingEXODUS (Chooramun et al., 2012, Galea et al.,
2008, Galea et al., 2011, Pretorius et al., 2013 and Siddiqui and Gwynne, 2012)
software and consists of three components; urbanEXODUS, webEXODUS and the
EXODUS engine as shown in Figure 1. Each of these are briefly described in the
following section.
Figure 1: Diagram showing the usage of urban and webEXODUS
3.1 urbanEXODUS
A desktop interface, urbanEXODUS, was developed to enable the EXODUS engine to
easily represent large scale urban spaces. It is capable of receiving geospatial vector
data from an OSM XML (Neis et al., 2012) file and converting it into a virtual
representation of space that is suitable for modelling pedestrian evacuation. It is also
capable of producing simulation output in the form of shapefiles that can be published
to GIS servers thus aiding visualisation of simulation output on web based GIS
interfaces. The main use of urbanEXODUS is pre-incident during the planning and
preparation phase of a possible large scale evacuation by simulating various what-if
evacuation scenarios.
While MATSIM only models the street network from an OSM XML file,
urbanEXODUS models additional real world spatial objects that pedestrians may make
use of during an evacuation including buildings, open spaces (public spaces, car parks,
piers, parks, etc). Once the virtual environment is modelled within urbanEXODUS the
user can then specify the population, hazards and procedural data, to define an
evacuation scenario. In addition to generating an xml output file providing details of
the evacuation simulation results, urbanEXODUS also generates the population density
contours in a geospatial data format such as shapefiles (Xia and Wei, 2008) so these
can be overlayed on baselayers such as Google Maps, OpenStreetMaps or Bing Maps
on a browser.
3.2 webEXODUS
A portable GIS based web interface, webEXODUS, was developed to be utilised during
an incident. It is connected to the evacuation simulation server over the Internet and can
be accessed remotely via browsers on desktop PCs, tablets or smartphones. It can also
be used during the planning and preparation phase similar to urbanEXODUS. However,
it is primarily intended for use during an incident by strategic command in the command
and control centre providing a generic data model to describe a variety of evacuation
simulations and run these simulations on the remote evacuation server. By enabling the
current tactical situation to be represented in the evacuation simulation and providing a
set of tools to rapidly analyse the results the system assists strategic commanders decide
which evacuation procedure to employ based on the situation as it unfolds.
webEXODUS requires the EXODUS geometry to have been prepared prior to the
incident enabling end users to simply specify the evacuation scenario(s) they wish to
consider during the actual incident. The evacuation scenarios can include expected (e.g.
generated from census data) or observed (e.g. generated by local authorities) population
levels and distributions, hazard locations (e.g. generated from model predictions or
observational data from operatives in the field) and evacuation procedures.
webEXODUS is connected to the EXODUS engine over the Internet and can initiate
new simulations remotely. After completion of the simulation, it is possible to view the
evacuation progress as population density contours on a standard browser.
webEXODUS has been integrated with the IDIRA disaster management system.
During an incident, it allows users in the command and control centre to publish the
results of an evacuation scenario that is most representative of the ensuing event to the
IDIRA COP, a component of the IDIRA system, that provides the incident commanders
in the field a summary of the myriad components involved in the rescue and relief
efforts following a large scale disaster. This information assists the commanders in the
field to make informed decisions in coordinating the evacuation of people during
disasters.
3.3 EXODUS Engine
The EXODUS engine (Chooramun et al., 2012, Galea et al., 2008, Galea et al., 2011,
Pretorius et al., 2013 and Siddiqui and Gwynne, 2012) receives simulation inputs from
urbanEXODUS or webEXODUS, performs the simulations and returns the results.
Within the EXODUS engine agents are represented as individuals with each agent
being defined by a set of attributes. The attributes broadly fall into four categories:
physical (such as age, gender, agility, mobility), psychological (such as patience, drive,
response time), experiential (such as distance travelled, time travelled, time waiting in
congestion) and physiological (such as respiration rate, impact of narcotic and irritant
gases, impact of heat). These attributes have the dual purpose of defining all occupants
as individuals while allowing their progress and condition during the evacuation to be
tracked. Agents can also be provided with itineraries which can modify their behaviour
prior to and during the evacuation process. For example, an agent may decide to collect
their children from school prior to starting to evacuate, or may need to undertake a
number of tasks in their dwelling prior to starting to evacuate.
The EXODUS engine is a hybrid model capable of utilising coarse node, fine node and
continuous spatial representations (Chooramun et al., 2012). The coarse and fine node
modes of the EXODUS engine have been utilised with the former providing quick
preliminary results at a faster than real time rate and the latter providing more refined
results but requiring longer run times. The coarse node model employs ‘location
estimation’ techniques by estimating the location and spread of agents within the
region. This is to partially compensate for inherent uncertainties associated with pure
coarse node modelling thus improving the accuracy of the movement model and hence
the overall predictions. It also allows for the calculation of population densities within
sub-regions of the coarse nodes. The same technique is also utilised to record the
number of people in 6m X 6m grid cells which are used to produce a heat map showing
population density contours. The coarse node model also employs varying walk speeds
depending on whether the agent is traversing paved paths or off road paths. Junctions
such as road junctions are modelled as special coarse nodes which can represent multi-
directional flow.
3.4 System Overview
An overview of the utilisation of the three components - urbanEXODUS, webEXODUS
and EXODUS engine during the preparation/planning and emergency responses phases
of disaster management is provided in Figure 2. Sometime before an incident, the
spatial data and hazard simulation data are provided as inputs to urbanEXODUS. The
spatial data is currently in the form of an OSM XML file. The OSM data includes the
coordinates of the geospatial objects as well as their description. For example the OSM
data includes the co-ordinates of an object and includes a description identifying the
type of object such as a building, road or park. urbanEXODUS reads the OSM data and
builds a virtual representation of space to model pedestrian movement in it. The KML
fire perimeter output from the fire spread model, Prometheus (Tymstra et al., 2010) was
also provided as an input to urbanEXODUS. This helped identify the times at which
critical areas in the modelled area became untenable due to fire.
During the planning and preparation phase, end users can then configure an evacuation
scenario by specifying population and procedural data. Population data mainly includes
the number of people in the region, their attributes and manner of their distribution.
Procedural data includes the manner in which the population is to be alerted, population
response times, identification of personal itineraries for the population and routes taken.
The EXODUS engine receives the scenario input from urbanEXODUS and runs the
simulation. During simulation, the EXODUS engine stores the population density
contours as the simulation progresses at pre-defined time-steps in the form of
Shapefiles. Upon completion of the simulation, the simulation results are transferred to
urbanEXODUS. The key results include the overall evacuation times, individual
evacuation times of each agent in the simulation along with other results such as
congestion times and safety margins. The Shapefiles are stored on a GIS server thus
allowing the simulation progress to be overlaid on basemaps using webEXODUS.
During the emergency response phase, webEXODUS could be utilised by strategic
level incident commanders to view the results of the simulations performed during the
planning and preparation phase thus assisting them to decide what evacuation procedure
to employ. However, not all eventualities can be tested prior to an incident and hence
incident commanders can also modify scenarios using the latest tactical information
reflecting the current situation and rerun the simulation. The web interface of
webEXODUS has been simplified so it can be utilised by those who are unfamiliar with
evacuation simulation tools. After analysing different evacuation procedures, the
incident commanders can publish the most suitable evacuation procedure to a disaster
management system or COP (Prinz, 2014) which provides a Common Operational
Picture in real-time, sharing disaster related information from all parties on strategic,
tactical and operational levels. The evacuation simulation results published on the COP
assists the tactical and operational commanders to manage the tactical implementation
of the evacuation process by implementing a set of evacuation actions.
Figure 2: System overview depicting the use of urbanEXODUS during
planning/preparation and webEXODUS during emergency response phases of a
large scale incident.
4 urbanEXODUS Application Example – Swinley Forest Fire
4.1 Initial conditions
The Swinley forest fire was contained within a 110 hectare region of the Swinley forest
(located in southeast England), directly threatening built-up areas. In this example the
population within the built-up area is hypothetical and used purely for demonstration
purposes. The threatened area consisted of, the Transport Research Laboratory (TRL),
consisting of approximately 800 people of mixed working age, a large business estate
(BE), consisting of approximately 200 people of mixed working age and predominantly
young, a pub, consisting of approximately 200 people of mixed age with a number of
elderly and 10% having some form of minor moving disability, and five residential
dwellings (RD), each consisting of a family of two adults and two children. The fire
also threatened near-by towns of Bracknell and Crowthorne, the Broadmoor High
Security Hospital and the Royal Military Academy Sandhurst. The fire, the largest in
Berkshire’s history, was attended at one point by 55 appliances from 12 brigades.
Fortunately, a road running along the built-up area (B3348) acted as a fire break
preventing the fire from spreading to the built-up area. Following this fire considerable
analysis was undertaken to determine the impact on the built areas had the wind
changed (Gazzard, 2014). One scenario considered a wind change of 900 to the North
West. This was modelled using the Prometheus fire spread modelling tool (Smith,
2014). If these conditions had occurred during the fire it would have prompted the
evacuation of the nearby built-up area (see Figure 3).
An overview of the area modelled in EXODUS is shown in Figure 3 (total area, 1.58
km2). The fire simulations are based on real meteorological data (e.g. fuel maps and
ignition locations/times) as recorded on 2 May 2011, however, with slightly different
wind conditions (speed and direction). The output shows the simulated spread of the
fire for over seven hours using 30-minute intervals. The residents of the indicated
buildings are evacuated to the assembly area at the Great Hollands recreation ground,
located on the north east side of the built-up area. All the routes leading to the assembly
area used by the agents in the simulations are shown in Figure 3.
Figure 3: The simulated area in Bracknell which was affected by the wildfire of
May 02, 2011 with the simulated hypothetical fire front at 7 hrs after fire
initiation (in 30-minute intervals).
4.2 EVACUATION SCENARIOS
The evacuation process is undertaken on foot as it is assumed that the roads need
to be kept free for use by emergency vehicles and road conditions are considered
hazardous for vehicles due to reduced visibility resulting from smoke. A total of 1220
agents were distributed in the directly threatened locations (pub, TRL, BE and RD).
During a wildfire affecting large areas it is essential for incident commanders to know
which areas to evacuate first (Cova et al., 2005 and Pultar et al., 2009). To do this they
need to know when the fire is likely to directly threaten the occupied area, when the
fire is likely to threaten proposed evacuation routes, how long is it likely to take to clear
the threatened areas and how long it will take for the evacuating population to pass
through the at risk evacuation routes. Key locations on the evacuation routes, labelled
A to G, are depicted in Figure 3. These represent ‘critical locations’ which the
evacuating population must pass to be considered safe from the approaching fire. These
locations were determined by analysing both the fire progression within the fire
simulation and the movement of the population within the evacuation simulation.
However, rather than the fire front we make use of the hazard front. The hazard front
is the region ahead of the fire front that is considered to be untenable. At this point the
smoke concentration, visibility and temperature are considered untenable. Ideally, the
forest fire model should provide information relating not only to the location of the fire
front, but also the position, optical density and composition of the smoke front hence
enabling the identification of the hazard front.
Unfortunately, current forest fire models, including Prometheus, do not have this
capability and so a 250 m region ahead of the fire front is assumed to mark the
hazardous extent of the fire smoke, here called the hazard front. Thus by knowing the
location of the fire front at any point in time (as determined by the fire model), the
location of the hazard front, ahead of the fire front, can be determined. Where the
hazard front crosses an evacuation route a critical region is identified (see Figure 4).
For the evacuation route to be considered safe, the last of the agents using that route
must have passed through the critical region and hence be beyond a critical location
defining the boundary of the critical region before the hazard front reaches the critical
location. The safety margin associated with the compromised route is the difference in
time between the time for the predicted hazard front to reach the critical location and
the predicted time for the last person to pass through the critical location.
This concept is demonstrated in Figure 4. Here the hazard front crosses the evacuation
route at time 𝑡6. The location of the hazard front at this time is identified by
𝐿𝑡61 and 𝐿𝑡6
2 (critical locations) with the critical region being defined as the region on
the evacuation route between these critical locations (in this example, 𝐿𝑡61 <
𝑐𝑟𝑖𝑡𝑖𝑐𝑎𝑙 𝑟𝑒𝑔𝑖𝑜𝑛 < 𝐿𝑡62 ). The last person in the population moving along the
evacuation route has location at time 𝑡𝑖 defined by 𝐿𝑡𝑖𝑃 . Thus for the population to be
considered safe, the last person must be beyond the critical region defined by the trailing
critical location in the direction of travel and so in this example 𝐿𝑡6𝑃 > 𝐿𝑡6
2 . The safety
margin is then defined as the difference between the time at which the hazard front is
at the critical location, in this example given by 𝐿𝑡62 (𝑡6), and the time at which the last
person was at the critical location, in this example given by (𝑡4), i.e.
𝑡6 − 𝑡4, where 𝑡6 is determined from the fire simulation and 𝑡4 is determined from the
evacuation simulation.
As the fire progresses, the hazard front may follow the evacuating population creating
a sequence of critical locations and associated safety margins for a given portion of
evacuation route. The safety margin for the population on the affected route is then
taken as the minimum of the sequence of safety margins and the associated critical
location is identified. Ideally, this sequence would be determined automatically by the
coupled fire and evacuation simulation environment which would then determine the
minimum safety margin and associated key critical location. However, in the work
presented here the most critical locations for each route and each population (generating
the minimum safety margins) are manually estimated by viewing both the fire and
evacuation simulation. Thus the reported safety margins can only be considered
approximate.
Figure 4: Defining critical region on an evacuation route
Four scenarios are investigated; they consist of a base case in which it is assumed that
all evacuation routes are available and three other cases in which various evacuation
routes are considered non-tenable. Each case assumes the same population composition
and distribution and the same fire conditions. An overview of the four scenarios is
provided in Table 1. Within each figure in the table, the arrows indicate the direction
of travel of the agents from the buildings. The dashed lines with a cross denote loss of
routes.
Table 1: Overview of evacuation scenarios.
Scenario Diagram Description
1
Base case scenario assuming that all routes are
available. Hence all agents take the shortest route
to the assembly location. The critical locations A
(TRL), B (Pub), C, G (RD) and E (BE) represent
the locations that each identified group of agents
must reach prior to the arrival of the hazard front.
2
Assumes the path out of TRL connecting to the
Nine Mile Road (B3430) is closed/unusable due
to use by emergency vehicles. The agents in TRL
are thus forced to take a longer route through the
BE. The critical location for the agents in TRL is
now F instead of A.
3
Assumes that a part of the Bracknell road is
unsafe to use forcing the agents in the residential
dwellings to take a slightly longer route via Old
Wokingham Road - Nine Mile Road to the
shelter location. In this case, the critical location
for the agents in the RD is D instead of C. This
part of the Bracknell road is the first route to be
affected by the fire and hence this is an important
scenario.
4
This scenario assumes that the route losses from
Scenarios 2 and 3 are merged in this scenario.
All the scenarios have the same initial conditions in terms of the number, nature, and
distribution of agents, their response times and the location of the assembly area which
is at the Great Hollands Recreation centre located in the north east. The four scenarios
differ in terms of the routes taken by the evacuating population. The routes utilised by
the agents in the four scenarios and the critical locations for the various establishments
are shown in Table 1. The critical locations A to G differ for each scenario depending
upon the routes taken by the agents. For example, in Scenario 1, the agents located
within the TRL building are considered to be safe after clearing critical location A.
However, in Scenario 2 a different path is adopted by the TRL agents and so the critical
location for these agents is now location F.
Within each scenario the agent response times were specified as follows:
Pub: 30 - 60 s, proprietor alerted by phone call from police at t = 0 s.
TRL: 60 - 120 s, building managers alerted by phone call from police at t = 0
s.
BE: 60 - 120 s, building managers alerted by phone call from police at t = 0 s.
RD: 10 to 16 min, alerted by the police door knock. Each dwelling is visited in
turn by a police patrol that informs residents to evacuate. The police reach the
first house at t = 300 s. The second house is visited 60 s later and so on. Each
household is arbitrarily assigned a 5 min response time once alerted. It is
assumed that the residential dwellings have been pre-warned of the possible
need to evacuate and so have prepared for the evacuation.
Occupant response times for these types of situations are not known and so the values
used in these simulations are intended for demonstration purposes. In reality it is likely
that people would have much longer response times, especially for RD. Furthermore, it
is possible that some of the occupants of the RD may chose not to evacuate but to defend
their properties. In addition, the start of the evacuation (t = 0 s) occurs at the time of
fire initiation. This is unlikely to be the case in reality as it will take some time for the
fire to grow to a size that can be detected and then for the fire to develop to the point
that it may be considered a threat. However for simplicity in this demonstration the
evacuation process is considered to start at the start of the fire. The walking speeds of
the agents in the simulations is randomly allocated between 0.85 m/s to 1.1 m/s. Note
that these speeds are less than the normal defaults used in buildingEXODUS which are
usually between 1.2 m/s and 1.5 m/s. In reality the walking speed of people would
depend on a range of factors such as age, gender, disability, nature of group, fatigue,
etc. These simplifications should be taken into consideration when assessing the results
presented.
5 Results and Discussion
The evacuation simulation results for the four scenarios are presented in this section.
Each scenario was run ten times and the values referred to in this section represent the
average of these ten simulations. The arrival times and safety margin graphs for each
scenario are presented and discussed in this section. The arrival times graph shows the
times at which agents arrive at the assembly location. The safety margins graph shows
the time for the fire front to be within 250 m of a critical location, the time for the last
person to pass through the critical region and the safety margin, which is the difference
between these two times. It is noted that all the simulations are performed using an
Intel Xeon E5-1620 computer with a clock speed of 3.6 GHz and 64 GB of RAM.
5.1 Scenario 1
5.1.1 Arrival times
The arrival times of the entire population at the assembly location is shown in Figure
5. The agents in the pub are closest to the assembly location, have a short response time,
and hence are the first to arrive at 13 min. All 200 agents from the pub reach the
assembly location within 19 min from the start of the simulation. After the pub, the
agents from TRL are the nearest and start to arrive in the assembly area at around 22
min. The 800 agents from TRL assemble within 35 min. The first agents from the BE
arrive at around the same time as the last agents from TRL and start assembling at 35
min. The 200 agents from the BE assemble by 52 min after the start of simulation. The
agents from the RD are the last to arrive as they have much higher response times and
distance to travel compared to the rest of the population. The 20 agents in the RD and
hence the entire population assembles in 1 hr 12 min 45 s (72.75 min). The EXODUS
engine provided the simulation results in 3 min 22 s. It was thus possible to run the
simulation 21.4 times faster than real time.
Figure 5: Arrival times at the assembly location for Scenario 1
5.1.2 Safety margins
The time for the agents to clear the relevant critical locations, the time for non-tenable
conditions to develop at the critical locations (i.e. the time for the hazard front to reach
the critical location) and the safety margin (i.e. the difference between the time for non-
tenable conditions to develop at the critical location and time for the agents to clear the
critical locations) for Scenario 1 is shown in Figure 6. The agents in the pub are quick
to clear critical location B, doing so in 4 min 28 s. The fire approaches critical location
B at 1 hr 31 min. The agents in the pub therefore have a safety margin of 1 hr 26 min
30s. The safety margin is intended to compensate for a number of uncertainties within
the evacuation modelling including for example, delays in notifying the population,
delays in population response, slower walking speeds, etc. For example, the pub
residents could be warned of the need to start their evacuation up to 1 hr 26 min 30s
later than assumed within the simulation and they would still be able to reach safety.
The agents from TRL and the BE have the largest safety margins (more than 4 hr) as
they are quick to clear their relevant critical location and the fire is predicted to take a
longer time to affect their routes. In this scenario, the agents in the RD have the smallest
safety margin of just 12 min which is primarily due to them requiring a relatively long
time (48 min) to clear the relevant critical location G. The hazard front is also quick to
approach this critical location at 1 hr. At the other critical location C, located further
along the evacuation route, the agents from the RD have a safety margin of 37 min 6 s
which is mainly due to the hazard front taking a longer time to reach this location at 1hr
and 31 min. Clearly the agents from the RD have to clear both critical locations (C and
G), but the critical location with the smallest safety margin is critical location G and so
this defines the critical safety margin. Though the hazard front also approaches critical
location B in approximately the same period of time, the agents in the pub clear this
critical location very quickly (less than 5 min) and hence this is not considered a great
threat for the pub occupants. If the people in the RD are to take the Bracknell road to
reach the assembly location it is essential that they are warned as quickly as possible.
Thus the police doing the door knock must not lose time in road congestion getting to
the RD, once they arrive they must spend very little time explaining the situation and
the residents must then respond to the warning quickly as they only have a safely margin
of 12 min. Given how tight the safety margins are for the RD in this scenario, the
evacuation strategy adopted in this scenario is not considered viable.
Figure 6: Evacuation and fire timelines for the different establishments in
Scenario 1
5.2 Scenario 2
5.2.1 Arrival times
Similar to Scenario 1, the agents in the pub assemble within 19 min (see Figure 7) as
there is no change to their evacuation route. In Scenario 2 agents from TRL are forced
to take a longer path via the BE as shown in Table 1. Thus in this scenario the agents
in TRL have a longer route to the assembly location compared to the agents from BE.
The agents from the BE are thus the next to start arriving at the assembly location at 34
min. Due to a distribution of walk speeds between 0.85m/s and 1.1 m/s the fastest agents
from TRL assemble before the slowest agents from BE. The agents from BE and TRL
assemble within 1 hr 7 min. The last agents to arrive are again the agents from the RD.
The first agents from the RD begin to arrive at around the same time as the last agents
from TRL. The last of the agents from the RD and hence the overall assembly time for
this scenario is 1 hr 11 min which is very similar to Scenario 1. The total evacuation
times have not changed significantly in these two scenarios as the agents from the RD
are always the last to arrive and their route has not changed. However, 800 agents from
TRL take a longer route (3 km) to the assembly location in Scenario 2 which is twice
the distance of the route taken in Scenario 1 (1.5 km). Furthermore, the average level
of congestion experienced by the agents from TRL (measured by the average
Cumulative Wait Time or CWT) is 84% greater than the level of congestion they
experienced in Scenario 2. The average time a TRL agent spent in reaching the
assembly location (measured by the Personal Elapsed Time or PET) has increased by
93% in Scenario 2. This indicates that even though the overall evacuation times are
similar in Scenarios 1 and 2 the two scenarios differ significantly in terms of the
evacuation dynamics experienced by the various sub-populations. The EXODUS
engine provided the simulation results in 3 min 38 s. It was thus possible to run the
simulation 19.5 times faster than real time.
Figure 7: Arrival times at the assembly location for Scenario 2
5.2.2 Safety margins
The safety margins for the agents in the various buildings in Scenario 2 are shown in
Figure 8. The agents in the pub and the RD have similar safety margins as in Scenario
1. The agents in TRL now have two critical locations to clear – critical locations F and
E. Clearing the first critical location F indicates that they have cleared the TRL site and
hence are safe in the event of the fire approaching the TRL site. Critical location F is
closer to the approaching hazard front than critical location A, thus the time for the fire
hazards to approach critical location F is shorter than the time for the fire hazards to
approach critical location A. This is the main reason for the 29 min decrease in the
safety margin available for the agents from TRL in Scenario 2 compared to Scenario 1.
The TRL agents thus have a 4 hour 21 min 18 s safety margin at critical location A. In
this scenario the TRL agents have to take a similar path to the assembly location as the
agents from BE. Hence they need to also clear the critical location E along with the BE
agents. The BE agents are the first to clear the location E at 13 min 18 s giving them a
safety margin of 6 hr 18 min 42 s. The first of the TRL agents to arrive at location E do
so at14 min which is approximately 1 min after the last of the BE agents have left the
region. Thus there is no interaction between the agents from the BE and TRL at location
E. The agents from TRL clear location E at 26 min 36 s giving them a safety margin of
6 hr 5 min 24 s. Since the TRL agents have to clear two critical locations A and E the
lower safety margin 4 hr 21 min 18 s is considered as their critical safety margin in this
scenario.
Figure 8: Evacuation and fire timelines for the different establishments in
Scenario 2
5.3 Scenario 3
5.3.1 Arrival times
The difference between Scenario 1 and Scenario 3 is the loss of a part of Bracknell
Road forcing the agents in the RD to take a slightly longer route. There are 20 agents
from the RD who are the last to reach the assembly location. Therefore, the arrival times
graph for Scenario 3 (see Figure 9) is similar to that of Scenario 1 (see Figure 5) except
for a small difference towards the end of the evacuation which is a result of the RD
agents taking a slightly longer time to reach the assembly region. The assembly time
for the agents from the RD and hence the overall assembly time for the entire population
in this scenario is 1 hr 26 min which is greater than Scenario 1 by 15 min. The overall
assembly time for this scenario is 1 hr 26 min 14 s. The EXODUS engine provided the
simulation results in 3 min 55 s. It was thus possible to run the simulation 22 times
faster than real time.
Figure 9: Arrival times at the assembly location for Scenario 3
5.3.2 Safety margins
The safety margins for the agents in the various buildings in Scenario 3 are shown in
Figure 10. Although Scenario 3 produces the longest overall evacuation time of the
three scenarios, the agents in the RD have a higher safety margin, 2 hr 45 min compared
to only 37 min in Scenarios 1 and 2. This is because the critical location for the RD on
this alternative route, location D, is further away from the approaching fire front than
critical location C. The agents in the RD are also slightly closer to critical location D
than critical location C and thus pass critical location D sooner than they would have
passed critical location C. The safety margins for TRL, BE and the Pub remain the
same as those for Scenario 1.
Figure 10: Evacuation and fire timelines for the different establishments in
Scenario 3
5.4 Scenario 4
5.4.1 Arrival times
The arrival times of the agents from the various buildings at the assembly location are
shown in Figure 11. The overall shape of the arrival curve is similar to Scenario 2 as in
both these scenarios the TRL agents take a slightly longer route to the assembly station.
However, the total assembly time in this scenario is similar to that in Scenario 3 as the
last agents to assemble in both Scenarios 3 and 4 are the agents from the RD and in
both scenarios these agents are forced to take a slightly longer route as part of the
Bracknell Road is assumed to be unusable. The overall assembly time for Scenario 4 is
1 hr 24 min 38 s. The EXODUS engine provided the simulation results in 4 min 12 s.
It was thus possible to run the simulation 20.2 times faster than real time.
Figure 11: Arrival times at the assembly location for Scenario 4
5.4.2 Safety Margins
The safety margins for the agents in the various buildings in Scenario 4 are shown in
Figure 12. The safety margin available for the agents in the pub is almost identical
across all four scenarios as their path has not been altered. The safety margin available
for the TRL agents is the same as that in Scenario 2 as in these two scenarios the agents
from TRL take the same route. Similar to Scenario 2 the TRL agents have a second
critical location, location E to clear along with the agents from the BE. The BE agents
clear this location first at 13 min 6 s and have a safety margin of 6hr 18 min 54s. There
is a gap of approximately 1 min between the times for the last BE to clear location E
and the time for the first TRL agent to reach it. The TRL agents clear location E at 26
min 42 s giving them a safety margin of 6 hr 5 min 18 s. The lower of the two safety
margins for the agents from TRL, which is 4 hr 21 min 18 s at location F, is hence more
critical in this scenario. The safety margin for the agents from the RD is approximately
the same as that for Scenario 3 as in these two scenarios the agents from the RD utilise
the same path.
Figure 12: Evacuation and fire timelines for the different establishments in
Scenario 4
5.5 Comparison of the four scenarios
5.5.1 Assembly times and distance travelled
The assembly times and distance travelled by the agents from the different locations in
each scenario is shown in Table 2. The agents from the pub are always the first to
assemble. The RD agents are always the last to assemble in all scenarios and thus
determine the overall assembly times for the entire population. A graph of the times at
which the agents arrive at the assembly locations for all the four scenarios is shown in
Figure 13.
The evacuation dynamics for Scenarios 1 and 3 are quite similar and the evacuation
dynamics of Scenarios 2 and 4 are quite similar, but both pairs exhibit quite different
evacuation dynamics to each other. The evacuation dynamics in Scenarios 1 and 3 are
similar as the only difference between the two scenarios is that the agents from the RD
take a slightly longer route in Scenario 3 (2.7 km versus 3.0 km) compared to Scenario
1. In these scenarios we note that there is a rapid build-up of people in the assembly
area after approximately 21 min. This is due to the arrival of people from TRL. While
the evacuation dynamics in these two scenarios are broadly similar, there is a large
difference (14 min or 20%) in the overall assembly times between these two scenarios.
This is a result of the agents from the RD travelling an additional 0.3 km to reach the
assembly location in Scenario 3.
The evacuation dynamics in Scenarios 2 and 4 are quite similar as again, the only
difference between these two scenarios is that the agents from the RD take a slightly
longer route in Scenario 4 (2.7 km versus 3.0 km) compared to Scenario 2. In these
scenarios we note that there is a slower build-up of people in the assembly area
compared to that in Scenarios 1 and 3. This difference is due to the occupants from
TRL taking a longer route to the assembly location (3.0 km versus 1.5 km) in Scenarios
2 and 4 compared to Scenarios 1 and 2. This has a significant impact on the
development of the arrival times. Indeed, as noted in Figure 3, after the arrival of the
last occupants from the pub into the assembly area there is a period of about 16 min
when no further agents arrive in the assembly location. While Scenarios 2 and 4 have
broadly similar evacuation dynamics, there is again a large difference (14 min or 20%)
in overall assembly times between these two scenarios. As with Scenarios 1 and 3, this
is due to the slightly longer evacuation route taken by the occupants of the RD in
Scenario 4 compared to Scenario 2.
The overall assembly times in Scenarios 1 and 2 are similar as are those in Scenarios 3
and 4. This is because the final assembly times are determined by the arrival times of
the RD occupants. In Scenarios 1 and 2 the occupants of the RD follow the same route
(most direct route) while in Scenarios 3 and 4 they follow the slightly longer route.
Hence the overall assembly times for Scenarios 1 and 2 are slightly shorter (71 min or
16%) compared to those for Scenarios 3 and 4 (85 min).
In all four scenarios the agents in the pub start arriving at the assembly location 13 min
after the call to evacuate. All 200 agents from the pub assemble after 20 min. The agents
in the pub take the same routes (0.8 km), have the same response times and do not share
their escape route with other agents for all the four scenarios. This explains why they
exhibit a uniform behaviour in all four scenarios. In Scenarios 1 and 3 the agents from
TRL take the next shortest route (1.5 km) to the assembly location. Hence in these
scenarios they arrive next and start to arrive at the assembly location after around 22
min. There is a 3 min gap between the last person arriving from the pub and the first
person arriving from TRL in Scenarios 1 and 3. The 800 agents from TRL travel 1.5
km and assemble within 36 min. The first agents from the BE arrive at the assembly
location about the same time as the last agents from TRL. The 200 agents from the BE
travel 2.3 km and assemble within 52 min. The agents in the RD are the next to arrive
and start to assemble after 61 min in Scenario 1 and 64 min in Scenario 3. This is
because the agents from the RD take a slightly longer route in Scenario 3 compared to
Scenario 1 – 0.3 km longer. All the agents from the RD are assembled after 1 hr 11 min
in Scenario 1 and 1 hr 26 min in Scenario 3.
Table 3: Times to assemble and distance travelled by the agents in various
establishments in the four scenarios
Scenarios Assembly times (HH:MM:SS) Distance travelled (km)
Pub TRL BE RD/ALL Pub TRL BE RD
1 00:18:38 00:35:26 00:51:36 01:11:32 0.8 1.5 2.3 2.7
2 00:18:37 01:07:05 00:52:09 01:10:26 0.8 3.0 2.3 2.7
3 00:18:42 00:35:35 00:51:16 01:26:01 0.8 1.5 2.3 3.0
4 00:18:40 01:07:25 00:52:31 01:24:25 0.8 3.0 2.3 3.0
Figure 13: Comparison of the arrival times at the assembly location. The times at
which proportions (50%, 80%, 95% and 100%) of the entire population
assemble is denoted by horizontal dashed lines.
5.5.2 Assembly Performance
The times at which various proportions (50%, 80%, 95% and 100%) of the population
arrive at the assembly location is provided in Table 4. As already noted there is a rapid
build-up of people in the assembly area in Scenarios 1 and 3 compared to Scenarios 2
and 4. This is due to the occupants from TRL taking the direct route to the assembly
station. From Table 3 we note that within Scenarios 1 and 3, not only do 50 % of the
population arrive at the assembly location in just under 31 min (compared to just over
55 min in Scenarios 2 and 4), but 95% of the total population have assembled in a
shorter time than the time taken for 50% of the population to assemble in Scenarios 2
and 4. Thus in Scenarios 1 and 3 it is essential that the assembly locations are staffed
and ready to handle the arrival of large numbers of people in a relatively short period
of time – 30 min from the start of the evacuation process.
Furthermore, in Scenarios 1 and 3 the arrival time curve has a long tail, the last 5% of
the population requiring between 26 min (in Scenario 1) and 41 min (in Scenario 3).
This is equivalent to the time required for the first 50% to arrive in the assembly in the
assembly location. Indeed, in Scenario 3, the last 5% of the population require almost
50% of the total assembly time.
Table 4: Times at which proportion of entire population assembled
Proportion
Assembled
50%
(610)
80%
(976)
95%
(1159)
100%
(1220)
Scenario 1 00:30:56 00:35:05 00:45:19 01:11:45
Scenario 2 00:55:22 01:02:30 01:05:58 01:10:54
Scenario 3 00:30:55 00:35:06 00:45:29 01:26:14
Scenario 4 00:55:20 01:02:40 01:06:11 01:24:38
5.5.3 Safety Margins
Presented in Table 5 are the safety margins associated with the various populations in
each scenario. The smallest safety margins are 11 min and 12 min for the occupants of
the RD in Scenarios 1 and 2 respectively and 87 min for the occupants of the pub in
Scenarios 3 and 4. The longest safety margins are over 360 min and are associated with
the population from the BE in all four scenarios.
The very short safety margins associated with Scenarios 1 and 2 are due to the
occupants from the RD taking the shortest route to the assembly location i.e. along
Bracknell Road, which inevitably takes them quite close to the fire start location and
directly in the path of fire spread. Given that the start of the evacuation process is
associated with the start of the fire, this means that there could only be a maximum of
11 min delay between the start of the fire and the notification of the population in the
RD before Scenarios 1 and 2 would become non-viable.
In Scenarios 3 and 4, the RD population take the longest route to the assembly location,
i.e. along Old Wokingham Road, which takes them effectively away from the hazard
front and hence gives them a much higher safety margin of over 165 min (associated
with clearing critical point D). Although the RD agents travel a longer route (0.3 km
more) via the Old Wokingham Road, it is the route that offers them maximum safety
and is the only viable route given the delay likely between starting the evacuation and
detecting the fire. The Pub occupants also have a relatively small safety margin
requiring them to start their evacuation within 86 min of fire initiation, regardless of
which scenario is considered. The agents in the other establishments (TRL and BE)
have more than four hour safety margin in all scenarios and hence can be considered to
be quite safe in terms of warning time and time needed to reach safety locations.
Table 5: Critical safety margins (min) and critical locations
Pub TRL BE RD
Scenario 1 87 (B) 290 (A) 379 (E) 12 (G)
Scenario 2 87 (B) 261 (F) 379 (E) 11 (G)
Scenario 3 87 (B) 290 (A) 379 (E) 166 (D)
Scenario 4 87 (B) 261 (F) 379 (E) 167 (D)
Thus of all four scenarios, Scenario 3 offers the greatest margin of safety for most
people, even though it results in the longest assembly time and the longest travel
distance for the population of the RD. Without the use of coupled fire and evacuation
modelling this result may not have been immediately apparent as it is somewhat counter
intuitive i.e. preferred option involves longest assembly time and greatest travel
distance for RD residents.
The safety margins also provide a basis on which to prioritise the warning sequence for
the occupants in the various establishments. In the preferred Scenario 3, the population
with the shortest safety margin are the occupants of the Pub followed by the occupants
of the RD and TRL and finally the BE. Thus if resources are limited, it would be
essential to warn the occupants of the Pub first followed by the RD.
5.5.4 Congestion on Evacuation Routes
Relatively small amounts of congestion was experienced on the evacuation routes in all
scenarios. Since the evacuation took place on roads which were at least 4 meters wide,
there was not much congestion experienced by the agents in any of the scenarios.
Furthermore, the agents from the different establishments reached the main exit routes
at slightly different times reducing the likelihood of generating high congestion levels.
The percentage of time spent by the agents in congestion compared to their overall
evacuation time for all scenarios is shown in Table 6.
Table 6: Percentage of the time agents spent in congestion compared to their
overall evacuation time
Scenario 1 (%) Scenario 2 (%) Scenario 3 (%) Scenario 4 (%)
TRL 7.2 5.7 7.2 5.7
BE 1.2 1.3 1.2 1.3
Pub 5.7 5.6 5.7 5.6
The agents from TRL experience the greatest levels of congestion for all scenarios with
Scenarios 1 and 3 producing the most severe conditions when they take the direct route
to the assembly area. However, when they use the path that takes them through the BE
the levels of congestion they encounter are lower due to the fact that their evacuation
takes longer and have longer time to spread on the evacuation route. However, in the
worst cases of Scenarios 1 and 3 they only waste 7.2% of their travel time in congestion.
The agents from the BE experience very low levels of congestion 1.2% for Scenarios 1
and 3 and 1.3% for Scenarios 2 and 4. This is due to the fact that the relatively fewer
agents (200 in total) are distributed in several buildings within the estate. Therefore,
there is little chance for congestion to build up along the evacuation route. Conversely,
the same number of agents that are located in the Pub experience higher levels of
congestion at 5.7% for Scenarios 1 and 3 and 5.6% for Scenarios 2 and 4. This is due
to these agents exiting from one small building and utilising a single short path (0.8
km) to the assembly area. The agents originating from the RD did not experience any
congestion in any of the scenarios.
5.6 Reviewing simulation results in webEXODUS during emergency response
The version of EXODUS used in the above analysis, which is equivalent to the planning
and preparation phase of an emergency was urbanEXODUS. During the emergency
response phase, the incident strategic commanders require a more mobile and easier to
use version of the evacuation simulation tool. In these applications the webEXODUS
interface can be used by operators with little or no knowledge of evacuation simulation
software to analyse evacuation procedures during a developing incident. webEXODUS
can be accessed over the internet allowing field commanders to specify simulation
inputs using a browser, run simulations on the EXODUS engine located on a remote
server and view the simulation results. A single aspect of the webEXODUS
functionality namely the ability to replay the simulation results on a web browser is
shown in Figure 14. The main purpose of this functionality is to allow the user to step
through the simulation to gain an understanding of how the pedestrian evacuation is
projected to progress during the incident. These figures show snapshots of the
simulation progress for one randomly selected simulation from the 10 repeat
simulations for Scenario 1 at 10 min, 30 min, 60 min and 91 min. Though it is possible
to show the evacuation simulation progress at 10 s time steps, these key time steps have
been chosen for discussion in this section as they correspond to time step used in the
fire model.
There are three main sections in the webEXODUS interface used to view the simulation
progress as a time stepped animation. The main body consists of the base map and the
population density overlays. The base map shown is the OpenStreetMap. However,
other base map layers can be selected including various versions of Google Maps. The
population density contour is seen as dots at the assembly location and at the
highlighted (circled) locations. The dots are cells measuring 36 m2 (6 X 6 meters) and
their colour represents a heat map based on the density of the agents contained in the
cell (e.g. hotter colour, red, for higher densities and cooler colour, blue, for lower
densities). The reason behind the selection of this size is based upon the assumption
that group behaviour is best observed at this size (Pretorius et al., 2013). The critical
locations A, B, C and E are shown as lines on the map labelled accordingly. These lines
or critical locations are interactive in that they can be clicked on to bring up a popup
showing the total number of agents that have cleared the location during the entire
simulation and the number of agents that have cleared them at the current time.
The snapshot in Figure 14a was captured at 10 min after fire ignition i.e. the start of the
evacuation. The fire data is provided at approximately 30 minute time steps and hence
there is no fire data available at this time. The footer of this interface contains the
controls to step through the evacuation simulation. The simulation time, the number of
agents that have reached the assembly location and number still evacuating is also
shown. If there are any predicted fatalities or agents trapped in the simulation due to a
loss of viable escape route they can also be depicted. In this simulation at 10 min, all
the agents in the Pub have cleared critical location B and 14 of them have already
assembled with the rest on their way to the assembly location. Agents from TRL and
BE are clearing safe locations A and E and are heading towards to the assembly
location. At this time only 14 agents have assembled, with 1206 yet to reach the
assembly location.
(a) Simulation progress at 10 min (b) Simulation progress at 30 min
(c) Simulation progress at 60 min (d) Hazard (smoke and fire) fronts at 91 min
Figure 14: webEXODUS user interface showing progress of evacuation and
hazard front at various times
The simulation progress at 30 min is shown in Figure 14b. The fire is still small and is
not considered to be a threat to the evacuating population. All the agents from the pub
have assembled. At this time, 491 agents from TRL have reached the assembly location
with the remainder of the TRL population walking along the South Road. The BE
agents are walking along the Nine Mile Ride Road. At this time 691 agents have
assembled, with 529 yet to reach the assembly location.
The simulation progress at 1 hr is shown in Error! Reference source not found.Figure
14c. Only a handful of agents have not reached the assembly location and are still
walking along the South Road. Though the fire front and projected smoke front has now
spread closer to the A3095 Road, all the agents are now clear of the advancing front
and can be considered to be safe. At this time only 1208 agents have assembled, with
only 12 yet to reach the assembly location.
The extent of the hazard front at 30, 60 and 91 min after ignition are shown in Figure
13d. The figure demonstrates how the critical locations and safety margins are
determined. In this work the hazard front is considered to be 250 m ahead of the fire
front. Thus by knowing the location of the fire front at a point in time, the location of
the hazard front, ahead of the fire front can be determined. At 30 min it is clear that the
fire front is far away from A3095, but the hazard smoke front is some 250 m ahead of
the fire front and so is within 45 m of the road. Thus while the hazard front is
approaching the A3095, the road is still considered passable at this time.
The last of the agents from the RD passes the point marked on the A3095 at 45 min.
Based on the fire model predictions, the hazard front crosses the A3095 sometime
between 30 min and 60 min. The location at which the hazard front crosses the A3095
could be considered a critical location and this is likely to occur sometime between 30
min and 60 min which may impact the agents from the RD. However, as the fire hazard
data was only available in 30 min time steps it was not possible to identify the critical
location on the A3095 and hence it was not possible to determine if the agents from the
RD are considered safe at 45 min. The last of the RD agents reaches critical location
G at 48 min 6sand critical location C at 54 min. At 60 min the hazard front has covered
part of the A3095 but the last of the RD agents have reached critical location C by 54
min and so have cleared the threatened part of the A3095. At 60 min, the hazard front
is predicted to have reached the vicinity of critical location G. As the last of the agents
from the RD cleared this point at 48 min, the safety factor for the RD determined at
critical location G is 12 min (60 min - 48 min). Similarly, at 91 min, the hazard front is
predicted to have reached the vicinity of critical location C. As the last of the agents
from the RD cleared this point at 54 min, the safety factor for the RD determined at
critical location C is 37 min (91 min - 54 min). Thus on this route, the hazard front
follows the RD agents during their evacuation. Assuming that the RD agents are safely
through the hazardous region when the hazard front reaches the road (possibly at around
45 min), the RD agents are able to stay ahead of the advancing hazard front but must
keep moving without reducing their travel speeds.
6 Limitations and further work
Urban scale evacuation simulation is still in its infancy and so the software presented
in this paper should not be considered a finished product but a prototype. More work
is required both in the development of the software tools and in understanding and
quantifying the behaviour and performance of people subjected to large-scale disasters.
Several important limitations of the work presented in this paper include:
Understanding and quantifying the response phase behaviour of people (Galea,
2009, Galea et al., 2013 and Galea et al., 2017) in large-scale urban evacuation
situations is critical to the validity of urban scale evacuation simulation models.
The response phase behaviour is likely to be different for different types of
hazardous situations such as wildfire, flood, earthquake and tsunami.
Furthermore, this is likely to be culturally and situation dependent (Galea et al.,
2013), with people in Hafia Israel potentially behaving differently to a wildfire
compared to people in Australia and people in a remote homestead in Australia
potentially reacting differently to people in a large urban conurbation in
Australia. Other factors such as experience of previous events, training,
severity of the perceived threat, prior warning, type of prior warning, social
bonds are also expected to exert an influence on response behaviour. The
authors are involved in a Horizon 2020 Marie Curie project called GEO-SAFE
(Fire Safety Engineering Group, 2016) which intends to collect data concerning
response phase behaviour of people subjected to the threat of wildfire in both
Europe and Australia which will provide some insight into this issue. Until
such data is collected, sensitivity analysis should be undertaken to estimate the
impact of response behaviours on model predictions.
Physical walking capabilities of pedestrians over long distances (possibly
several kilometres), the impact of fatigue and the impact of terrain type,
including steepness of grade and nature of surface e.g. paved, gravel, grassy,
etc, must be represented within urban scale evacuation simulation models. This
extends not just to the physical capabilities of the population but also how the
different terrain types can be identified and represented within the geometry of
the simulation model. Within the simulations presented in this paper, both
fatigue and the nature of the terrain was not represented. Thus the results
presented in this paper should be considered ‘optimistic’. In collaboration with
colleagues in France, ISMANS (Fire Safety Engineering Group 2015a) and
Italy, Corpo Nazionale Dei Vigili Del Fuoco (Vigili del Fuoco Comando
Perugia, 2016), the authors have started to collect human performance data to
address this need.
Large-scale evacuation situations are unlikely to be performed solely by
pedestrians on foot. Road vehicle traffic is a likely component of most urban
scale evacuation situations. Road vehicles are likely to be used as a means of
population evacuation and may even have an impact on the evacuation of
pedestrians. In addition, vehicles will be used to dispatch first responders such
as fire fighters to where they are needed and, as in the simulations presented in
this paper, provide a means of communicating the need to evacuate to scattered
members of the population by dispatching police officers. As part of the IDIRA
project (IDIRA end user comments, 2013) the evacuation simulation
environment presented in this paper was extensively tested by organisations
involved in disaster management who reported that the system was both user
friendly and a useful decision support tool for large scale disasters. The major
limitation that was identified was the omission of a representation of vehicles
within the simulation environment. Clearly, it is desirable to include within
urban scale evacuation simulation models a capability to not only represent
vehicles, as several models already do (Gulam and Rahim, 2004, Singh et al,
2016 and Halati et al., 2017), but to allow the interaction between vehicles and
pedestrians. To address this limitation, the authors have embarked on a research
project to include the representation of road vehicles within agent based
pedestrian evacuation simulation.
Clearly, if urban scale evacuation simulation is to be used not only for planning
but as part of a decision support tool for incident managers, the software must
be able to perform simulations much faster than real time. For the size of the
population represented in these simulations (1220 people) and for the size of
space represented (1.58 km2), the EXODUS engine was found to be capable of
simulating the evacuation about 20 times faster than real time using a standard
PC computer. This type of performance, and better, is required for populations
not measured in one or two thousand, but in the tens and hundreds of thousands
and areas measuring tens or hundreds of square kilometres. It is suggested that
further improvements in performance could be achieved using the parallel
implementation of the EXODUS engine (Grandison et al., 2017) and with the
use of more powerful PC computers.
Wildfire fire simulation tools such as Prometheus (Tymstra et al. 2010) and
Phoenix (Tolhurst et al. 2008) currently only provide an estimation of the
location of the fire front. They do not also provide information related to the
movement, optical density and composition of fire smoke. Knowledge of the
presence and concentration of fire smoke is vital for evacuation in wildfires as
smoke is likely to make evacuation routes non-tenable long before the flaming
fire front. Further development of wildfire fire simulation tools is therefore
required before the coupled fire-evacuation analysis can be truly useful as part
of the real time decision support tool for incident managers.
Calculation of the safety margins associated with each population and
evacuation route is currently a laborious manual process requiring the running
of many coupled fire and evacuation simulations to identify potential critical
locations and as such is prone to error. The process of identifying critical
locations and safety margins can be automated ensuring that all critical
locations associated with evacuation routes are identified and all safety factors
computed.
7 Conclusion
This paper presents a prototype software system that has been developed to evaluate
evacuation procedures during the preparation (urbanEXODUS) and emergency
response phases (webEXODUS) of a large scale incident such as a wildfire. The
software makes use of open source geospatial vector data (OSM XML) to generate a
virtual representation of the physical space through which pedestrian based evacuation
occurs. Free and open source web and GIS software is used to develop the web
interface which is used to establish a remote connection with the EXODUS engine and
integrate with a state-of-the-art web based disaster management system to produce a
Common Operating Picture for disaster managers. Furthermore, the software was
loosely coupled to the forest fire simulation software Prometheus enabling the potential
impact of the fire on the evacuation to be evaluated.
The simulation tool was demonstrated by simulating the evacuation of part of Bracknell
UK which was affected by the Swinley fires of 2011. The software, coupled with the
output from the Prometheus wildfire simulation software, was used to evaluate four
possible evacuation scenarios associated with hypothetical wind changes that could
have impacted the fire development during the actual fires. The analysis determined
the time required to evacuate the threatened population to a place of safety, the distances
travelled by the population, levels of congestion incurred during the evacuation and the
safety margins associated with each population centre in each scenario. The analysis
identified that two of the scenarios would result in very small safety margins associated
with the evacuation of occupants from threatened residential dwellings. The optimal
scenario, which resulted in maximum safety margins for all threatened sub-populations
was counter intuitive in that it resulted in the maximum assembly time and the longest
travel distances for the occupants of the residential dwellings. Use of the software also
confirmed the viability of the identified assembly area, prioritised the alerting of the at-
risk populations and assisted in prioritising the tasks of the emergency services.
The work demonstrates that the integration of fire simulation with evacuation
simulation enables the estimation of safety margins associated with alternative
evacuation strategies and thereby assists incident managers in planning phased
evacuation strategies. Most importantly, the quantification of safety margins associated
with evacuation choices assists incident managers to more reliably select the most
appropriate evacuation strategy for the evolving situation. The system can be
extrapolated to other emergencies such as those associated with floods and earthquakes
or manmade terrorist situations. Additional research is required to improve the
applicability and reliability of urban scale agent based evacuation simulation including,
improving our understanding and quantification of; response phase behaviour, walking
behaviour of pedestrians over long distances and different types of terrain, the
representation of vehicles and their interaction with pedestrians, the representation of
smoke in wildfire simulations and the speed and performance of large scale agent based
simulation models.
8 Acknowledgements
Project IDIRA (FP7 261726) was funded under the European Union Framework 7
Security mission dealing with Interoperability of data, systems, tools and equipment in
case of crisis (SEC-2010.4.2-1). The University of Greenwich authors acknowledge
the co-operation of their project partners in undertaking this work and in allowing the
project findings to be published. The authors are also indebted Louise Osborn,
Emergency Planner, Bracknell Forest Council for assistance in defining the scenarios.
9 References
BBC News, 2016. Israel fires: Tens of thousands flee as fires hit Haifa. [online]
Available at: <http://www.bbc.co.uk/news/world-middle-east-38088651> [Accessed
30 Dec 2016].
Bernardini, G., D'Orazio, M., Quagliarini, E. and Spalazzi, L., 2014. An Agent-based
Model for Earthquake Pedestrians’ Evacuation Simulation in Urban Scenarios.
Transportation Research Procedia, 2, pp.255-263,
http://dx.doi.org/10.1016/j.trpro.2014.09.050.
Careem, M., Bitner, D. and Silva, R.D., 2007. GIS integration in the Sahana Disaster
Management System. In: ISCRAM 2007 - Intelligent Human Computer Systems for
Crisis Response and Management. Netherlands, 13-16 May 2007. Netherlands: Bartel.
Chicago Tribune, 2016. More than 82,000 people flee Southern California wildfire.
[online] Available at: <http://www.chicagotribune.com/news/nationworld/ct-
california-wildfire-20160816-story.html> [Accessed 30 December 2016].
Chooramun, N., Lawrence, P.J. and Galea, E.R., 2012. An agent based evacuation
model utilising hybrid space discretisation. Safety Science, 50(8), pp.1685–1694,
http://dx.doi.org/10.1016/j.ssci.2011.12.022.
Cova, T.J., Dennison, P.E., Kim, T.H. and Moritz, M.A., 2005. Setting Wildfire
Evacuation Trigger Points Using Fire Spread Modeling and GIS. Transactions in GIS,
9(4), pp.603–617, http://dx.doi.org/10.1111/j.1467-9671.2005.00237.x.
Cova, T.J. and Johnson, J.P., 2002. Microsimulation of Neighborhood Evacuations in
the Urban–Wildland Interface. Environment and Planning A, 34(12), pp.2211-2229,
http://dx.doi.org/10.1068/a34251.
Crowhurst, E., 2015. Fire statistics: Great Britain April 2013 to March 2014. London:
Department for Communities and Local Government.
Currion, P., Silva, C.D. and Walle, B.V.D., 2007. Open source software for disaster
management. Communications of the ACM, 50(3), pp.61-65,
http://dx.doi.org/10.1145/1226736.1226768.
D’Orazio, M., Quagliarini, E., Bernardini, G. and Spalazzi, L., 2014. EPES -
Earthquake pedestrians’ evacuation simulator: A tool for predicting earthquake
pedestrians’ evacuation in urban outdoor scenarios. International Journal of Disaster
Risk Reduction, 10(A), pp.153–177, http://dx.doi.org/10.1016/j.ijdrr.2014.08.002
Drake, J.D. and Worsley, J.C., 2002. Practical PostgreSQL. O'Reilly Media, ISBN:
978-1565928466
Durst, D., Lämmel, G. and Klüpfel, H., 2014. Large-Scale Multi-modal Evacuation
Analysis with an Application to Hamburg. In: Weidmann, U., Kirsch, U.,
Schreckenberg, M (Eds). Pedestrian and Evacuation Dynamics 2012. Switzerland:
Springer International Publishing, pp.361–369, http://dx.doi.org/10.1007/978-3-319-
02447-9_30.
Fire Safety Engineering Group, 2015a. ISMANS inclined plane trials, France 6 March
2015. [online] Available at: <https://www.facebook.com/pg/Fire-Safety-Engineering-
Group-FSEG-
136784013003514/photos/?tab=album&album_id=1038243406190899> [Accessed
30 Dec 2016].
Fire Safety Engineering Group, 2015b. IDIRA EU FP7 project. [online] Available at:
<http://fseg.gre.ac.uk/fire/idira.html> [Accessed 06 Jan 2017].
Fire Safety Engineering Group, 2016. GEO-SAFE EU Horizon 2020 Marie Curie
project. [online] Available at: <http://fseg.gre.ac.uk/fire/geo-safe.html> [Accessed 30
Dec 2016].
Galea, E.R., 2009. Evacuation Response Phase Behaviour, Paper No. 09/IM/147,
October 2009, University of Greenwich. London: CMS Press. ISBN: 978-1-904521-
62-4.
Galea, E.R., Deere, S.J, Hopkin, C.G., and Xie, H., 2017. Evacuation Response
Behaviour of Occupants in a Large Theatre during a Live Performance. Fire and
Materials, pp. 1-26, http://dx.doi.org/10.1002/fam.2424.
Galea, E.R., Filippidis, L., Deere, S.J., and Sharp, G., 2013. Behaviour - Security -
Culture. Human behaviour in emergencies and disasters: A cross-cultural Investigation.
Silke Schmidt and Ed Galea (Eds), pp 131-215, 2013, Pabst Science Publishers. ISBN:
978-3-89967-867-3.
Galea, E.R., Hulse, L., Day, R., Siddiqui, A. and Sharp, G., 2011. The UK WTC 9/11
evacuation study: an overview of findings derived from first-hand interview data and
computer modelling. Fire and Materials, 36(5-6), pp.501-521,
http://dx.doi.org/10.1002/fam.1070.
Galea, E.R., Sharp, G. and Lawrence, P., 2008. Investigating the Representation of
Merging Behavior at the Floor Stair Interface in Computer Simulations of Multi-Floor
Building Evacuations. Journal of Fire Protection Engineering, 18(4), pp.291-316,
https://doi.org/10.1177/1042391508095092.
Gazzard, R., 2014. Summary of Forestry Commission analysis of IRS for Wildfire
evidence. [online] Available at:
<http://www.figuk.org.uk/conferences/2014/rob_gazzard.pdf> [Accessed March 20,
2015].
Gulam, H.B. and Rahim, F.B., 2004. A high fidelity traffic simulation model based on
cellular automata and car-following concepts. Transportation Research Part C:
Emerging Technologies, 12(1), pp 1-32, http://dx.doi.org/10.1016/j.trc.2002.05.001.
Grandison, A., Cavanagh, Y., Lawrence, P.J. and Galea, E.R., 2017. Increasing the
simulation performance of large-scale evacuations using parallel computing techniques
based on domain decomposition, Fire Technology, pp.1-40, http://dx.doi.org/
10.1007/s10694-016-0645-8.
Halati, A., Lieu, H. and Walker, S., 1997. Corsim - Corridor Traffic Simulation Model.
In: American Society of Civil Engineers, Traffic Congestion and Traffic Safety in the
21st Century: Challenges, Innovations, and Opportunities. New York: The National
Academies of Sciences, ISBN: 0784402434.
Helbing, D. and Johansson, A., 2013. Pedestrian, Crowd, and Evacuation Dynamics.
Encyclopedia of Complexity and Systems Science, 16, pp.6476-6495. New York:
Springer, http://dx.doi.org/10.1007/978-0-387-30440-3_382.
IDIRA end user comments, 2013. Confidential end user feedback from the IDIRA
Twist training meeting 27 October 2013, CNVVF camp base, Italy.
Incident Information System, 2016. Wildfire On The San Bernardino National Forest.
[online] Available at: <https://inciweb.nwcg.gov/incident/4962/> [Accessed 30
December 2016].
Klüpfel, H., 2014. Large Scale Multi-modal Simulation of Pedestrian Traffic.
Transportation Research Procedia, 2, pp.446-451,
http://dx.doi.org/10.1016/j.trpro.2014.09.058.
Lämmel, G., Grether, D. and Nagel, K., 2010. The representation and implementation
of time-dependent inundation in large-scale microscopic evacuation simulations.
Transportation Research Part C: Emerging Technologies, 18(1), pp.84–98,
http://dx.doi.org/10.1016/j.trc.2009.04.020.
Lämmel, G., Nagel, K., and Rieser, M., 2008. Bottlenecks and Congestion in
Evacuation Scenarios: a Microscopic Evacuation Simulation for Large-scale Disasters.
In: Bazzan, A.L., Klugl, F., Ossowski, S. (Eds). Proc. of 5th workshop on Agents in
traffic and transportation, at International Conference on Autonomous Agents and
Multiagent Systems, pp. 54-61.
Mordvintsev, A. S., Krzhizhanovskaya, V.V., Lees, M. H. and Sloot, P.M.A., 2012.
Simulation of City Evacuation Coupled to Flood Dynamics. In: Weidmann, U., Kirsch,
U., Schreckenberg, M (Eds.). Pedestrian and Evacuation Dynamics 2012. Switzerland:
Springer International Publishing, pp.485-499, http://dx.doi.org/10.1007/978-3-319-
02447-9_40.
Neis, P., Zielstra, D. and Zipf, A., 2012. The Street Network Evolution of
Crowdsourced Maps: OpenStreetMap in Germany 2007–2011. Future Internet, 4(1),
pp.1–21, http://dx.doi.org/10.3390/fi4010001.
Nishino, T., Tsuburaya, S., Himoto, K. and Tanaka, T., 2011. Validation of a Potential-
based Evacuation Model of City Residents in Post-earthquake Fire. In: R. D. Peacock,
E. D. Kuligowski, & J. D. Averill (Eds.). 2011. Pedestrian and Evacuation Dynamics.
Springer International Publishing. pp.399-406.
Obe, R. and Hsu, L., 2011. PostGIS in Action. Manning Publications, ISBN: 978-
1935182269.
Osaragi, T., Morisawa, T. and Oki, T., 2013. Simulation Model of Evacuation Behavior
Following a Large-Scale Earthquake that Takes into Account Various Attributes of
Residents and Transient Occupants. In: Weidmann, U., Kirsch, U. and Schreckenberg,
M, (Eds). Pedestrian and Evacuation Dynamics 2012. Switzerland: Springer
International Publishing, pp.361–369, http://dx.doi.org/10.1007/978-3-319-02447-
9_39.
Osborne, M.J. and Rubinstein, A., 1994. A Course in Game Theory. MIT press, ISBN
978-0262650403
Perez, A.S., 2012. OpenLayers Cookbook, Packt Publishing Ltd, ISBN 978-
1849517843
Pretorius, M., Gwynne, S. and Galea, E.R., 2013. Large crowd modelling: an analysis
of the Duisburg Love Parade disaster. Fire and Materials, 39(4), pp.301–322,
http://dx.doi.org/10.1002/fam.2214.
Prinz, R., 2014. Using Google Glass in Crisis & Disaster Management. [online]
Available at: <http://idira.eu/images/GlassCrisis.pdf> [Accessed 30 December 2016].
Pultar, E. Raubal, M. Cova, T.J. and Goodchild, M.F., 2009. Dynamic GIS Case
Studies: Wildfire Evacuation and Volunteered Geographic Information. Transactions
in GIS, 13, pp.85–104, http://dx.doi.org/10.1111/j.1467-9671.2009.01157.x.
Regional Municipality of Wood Buffalo, 2016. Regional Municipality of Wood Buffalo
2016 Wildfire Recovery Plan. [online] Available at:
<http://asset.rmwb.ca/files/RMWB-2016-Wildfire-Recovery-Plan.pdf?v1> [Accessed
30 December 2016].
Siddiqui, A. A. and Gwynne, S.M.V., 2012. Employing pedestrian observations in
engineering analysis. Safety Science, 50(3), pp.478–493,
http://dx.doi.org/10.1016/j.ssci.2011.10.011.
Silva, F.N.D. and Eglese, R.W., 2000. Integrating simulation modelling and GIS :
spatial decision support systems for evacuation planning. The Journal of the
Operational Research Society, 51(4), pp.423–430, http://www.jstor.org/stable/254169.
Singh, D., Padgham, L. and Perera, S., 2016. Bushfire BLOCKS for Modelling
Community Evacuation. In: Monoghan, M. (Eds), Australasian Simulation Congress.
Victoria, Australia, 26-29 September 2016, pp.509-514. Australia: Simulation
Australasia, ISBN: 978-0-9925960-2-6.
Smith, T., 2014. Private Communication, 11 August 2014, 16:09.
Teague, B., McLeod, R. and Pascoe, S., 2009. Interim report: 2009 Victorian Bushfires
Royal Commission. Victoria: Parliament of Victoria, ISBN 978-0-646-51830-5.
Teague, B., McLeod, R. and Pascoe, S., 2010. 2009 Victorian Bushfires Royal
[1Commission Final Report Summary. Victoria: Parliament of Victoria, ISBN 978-0-
9807408-1-3.
The Globe and Mail, 2016. [online] Available at:
<http://www.theglobeandmail.com/news/alberta/the-fort-mcmurray-disaster-read-the-
latest-weekend/article29930041/> [Accessed 25 Jan 2016].
The New York Times, 2016. As wildfire rage, Israel suggests arson and asks for foreign
help. [online] Available at:
<http://www.nytimes.com/2016/11/24/world/middleeast/israel-fires.html?_r=1>
[Accessed 30 Dec 2016].
Tolhurst, K.G., Shields, B. and Chong, D., 2008. PHOENIX: development and
application of a bushfire risk management tool. Australian Journal of Emergency
Management, 23(4), pp47–54, http://apo.org.au/node/48885.
Tymstra, C., Bryce, R.W., Wotton, B.M., Taylor, S.W. and Armitage, O.B., 2010.
Development and Structure of Prometheus : the Canadian Wildland Fire Growth
Simulation Model. Alberta: Natural Resources Canada, Canadian Forest Service,
Northern Forestry Centre.
Van de Walle, B. and Turoff, M., 2008. Decision support for emergency situations.
Information Systems and e-Business Management, 6(3), pp.295–316,
http://dx.doi.org/10.1007/s10257-008-0087-z.
Vigili del Fuoco Comando Perugia, 2016. AF3 Project People Evacuation Test 23-07-
2016 Valsorda Gualdo Tadino (Perugia). [online] Available at:
<https://youtu.be/64NDuO_WyaY> [Accessed 30 Dec 2016].
Xia, K. and Wei, C., 2008. Study on Real-Time Navigation Data Model Based on ESRI
Shapefile. In: The 2008 International Conference on Embedded Software and Systems
Symposia, pp.174–178. Sichuan, 29-31 July 2008. Los Alamitos: IEEE Computer
Society, http://dx.doi.org/10.1109/ICESS.Symposia.2008.57.